Buying memory for an AI data centre now works the way buying jet engines works. You commit years ahead, to a specification, for a workload that does not exist yet.
Ordinary server memory is a late decision. It is a standard part, available from several suppliers, and it can be added near the end of a build because nothing else depends on which box it came from. High-bandwidth memory removed that convenience. It has to be manufactured to a demanding specification, bonded into a package with a particular accelerator, and qualified as a unit, which means the decision migrates to the front of the process and stops being reversible. As the memory business discovered when it stopped behaving like a commodity, the buyer’s problem changed at the same time the seller’s did.
So an operator planning a facility has to choose a memory configuration before knowing what will run on it, and then design the processors, packaging, networking, power and cooling around that choice. The contract is signed against a forecast.
Both directions are exposed. Commit early and the supply is secured, but the architecture is fixed against workloads that may favour a different balance of capacity and bandwidth by the time the building opens. That shift does not require AI demand to disappear. More efficient software, smaller models, or a change in how work is distributed would be enough. Wait instead, and the risk inverts: processors, power and floor space available, and no memory to make them useful.
There is no version of this decision that is merely operational. It is a bet on what the work will look like in three years, made by people who have to order now, and it is being made simultaneously by everyone building at scale, which is its own complication.
The unresolved question is which discipline this market ends up rewarding. Securing capacity earliest, or keeping enough flexibility to still be right when the forecast turns out wrong.
